README.md

---
library_name: transformers
tags:
- radiology
- medical_imaging
- bone_age
- x_ray
- crop
license: apache-2.0
pipeline_tag: object-detection
base_model:
- timm/mobilenetv3_small_100.lamb_in1k
---

This model crops hand radiographs to better standardize the image input for bone age models. 
The model uses a lightweight `mobilenetv3_small_100` backbone and predicts normalized `xywh` coordinates.
![figure](images/fig.png)

The model was trained and validated using 12,592 pediatric hand radiographs from the [RSNA Pediatric Bone Age Challenge](https://www.rsna.org/rsnai/ai-image-challenge/rsna-pediatric-bone-age-challenge-2017) using an 80%/20% split. 
On single-fold validation, the model achieved mean absolute errors (normalized coordinates) of:
```
x: 0.0152
y: 0.0121
w: 0.0261
h: 0.0213
```

To use the model:
```
import cv2
import torch
from transformers import AutoModel

model = AutoModel.from_pretrained("ianpan/bone-age-crop", trust_remote_code=True)
model = model.eval()
img = cv2.imread(..., 0)
img_shape = torch.tensor([img.shape[:2]])
x = model.preprocess(img)
x = torch.from_numpy(x).unsqueeze(0).unsqueeze(0)
x = x.float()

# if you do not provide img_shape
# model will return normalized coordinates
with torch.inference_mode():
  coords = model(x, img_shape)

# only 1 sample in batch
coords = coords[0].numpy()
x, y, w, h = coords
# coords already rescaled with img_shape
cropped_img = img[y: y + h, x: x + w]
```

If you have `pydicom` installed, you can also load a DICOM image directly:
```
img = model.load_image_from_dicom(path_to_dicom)
```